Clip F to I conversion to the INT32 range - #10373

Fixes #10372.

Changes proposed in this pull request:

• Clip F to I conversion to the INT32 range in f2i, and map NaN to 0. The cast had no range check, so on x86 every value above the range, inf and NaN became -2147483648, and an I image at 2147483647 converted to F and back did too. These casts are undefined behavior in C, so the result also depended on the CPU.
• Add a test with values at and past both ends of the range, inf, -inf and NaN.

>>> Image.new("F", (1, 1), float("inf")).convert("I").getpixel((0, 0))
-2147483648

With this change, it is 2147483647, the same clipping clip32() in Filter.c does. NaN becomes 0, which is what F to L and F to I;16 give.